Sign In

RoXL

274
1.8k
33
0
Verified:
SafeTensor
Type
Checkpoint Trained
Stats
425
Uploaded
Dec 16, 2023
Base Model
SDXL 1.0
Training
Steps: 19,000
Epochs: 12
Hash
AutoV2
DF67355C65
0
0
0
0
0

RoXL is the SDXL evolution of my Rendered model, though with a lot more data and variety thrown in. While version 1.0 was focused on characters, and they still make up the bulk of the dataset, I've added a great deal of images to improve it's ability to handle a wide variety of concepts. Trained off of a mix of MOHAWK1.8, BrightProtoNuke1.2, BriXLv5, and RoXL1.0 (which was fine tuned off of Yamer's UnstableDiffusers V5.)

My main focuses, and (hopefully) the strengths of the model are:

  • A distinctive style far apart from the base SDXL experience

  • Expressive and natural feeling faces and poses

  • Variety in face/body shapes and types. If you're tired of everyone looking like lifeless dolls, give it a try!

  • A variety of scifi and fantasy themes and keywords. (good elf ears, scifi outfits, etc.)

  • Extra flexibility and variety in ethnicities

  • Extra control over hair, bodytypes, etc.

Recommended (tested) Settings:

  • Euler A(ncestral) Sampler, 20-30 steps. 7-9 CFG. For a painterly approach I haven't seen much of a benefit above 20 steps, but for detailed lineart you may need a little bit more.

  • Standard SDXL VAE

  • It seems to work well with most settings I have tried, so go hog wild. I've generally found it requires fewer steps than base SDXL to get great results.

The 2.0 release of the model is a huge leap from 1.0. The entire dataset was recaptioned (and roughly doubled in size) and a new model was trained off of a mix of some of my favorites.

Initial sample images have not been inpainted, but the full body ones have been upscaled (equivalent of hiresfix) to allow faces to generate at full quality. I highly recommend learning some inpainting for images that show full bodies, however, as they can be vastly improved. Guides can be found here and here.

Finally, Training SDXL models gets pricy. If you want to help me continue experimenting and improving models, please consider tossing me a few dollars on ko-fi.